pith. sign in

arxiv: 2604.08424 · v1 · submitted 2026-04-09 · 💻 cs.AI · cs.LG

On-board Telemetry Monitoring in Autonomous Satellites: Challenges and Opportunities

Lorenzo Capelli , Leandro de Souza Rosa , Maurizio De Tommasi , Livia Manovi , Andriy Enttsel , Mauro Mangia , Riccardo Rovatti , Ilaria Pinci
show 3 more authors
Carlo Ciancarelli Eleonora Mariotti Gianluca Furano
This is my paper

Pith reviewed 2026-05-10 17:37 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords explainable AIanomaly detectionsatellite telemetryfault detectionconvolutional autoencoderonboard monitoringreaction wheels
0
0 comments X

The pith

A framework derives low-dimensional peepholes from neural activations to add semantic interpretability to onboard satellite anomaly detection with little extra cost.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a method for making neural anomaly detectors more explainable in autonomous spacecraft. It extracts low-dimensional, semantically annotated encodings called peepholes from intermediate activations of a convolutional autoencoder trained on reaction-wheel telemetry. These peepholes produce interpretable indicators that support identification, localization, and bias detection of anomalies. The approach adds only marginal computational overhead, which supports its use in onboard fault detection, isolation, and recovery systems where both reliability and human-understandable explanations matter.

Core claim

The central claim is that peepholes derived from intermediate neural activations in a convolutional autoencoder applied to reaction-wheel telemetry yield low-dimensional, semantically annotated encodings that enable the identification and localization of anomalies while supporting bias detection and requiring only a marginal increase in computational resources for onboard feasibility.

What carries the argument

Peepholes: low-dimensional, semantically annotated encodings derived from intermediate neural activations of a convolutional autoencoder, which supply the interpretable indicators for anomaly analysis.

Load-bearing premise

That peepholes extracted from intermediate activations will reliably yield semantically meaningful and actionable indicators for anomaly identification and localization on actual reaction-wheel telemetry data.

What would settle it

A controlled evaluation on real satellite telemetry containing documented anomalies where the resulting peephole indicators fail to distinguish or semantically match the known fault types.

Figures

Figures reproduced from arXiv: 2604.08424 by Andriy Enttsel, Carlo Ciancarelli, Eleonora Mariotti, Gianluca Furano, Ilaria Pinci, Leandro de Souza Rosa, Livia Manovi, Lorenzo Capelli, Maurizio De Tommasi, Mauro Mangia, Riccardo Rovatti.

Figure 1
Figure 1. Figure 1: Block scheme of the proposed framework extracting [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Block diagram of the autoencoder producing anomaly score, including [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Anomaly identification when applied to all channels or to a single [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Confusion matrices for RW identification using Peephole for five [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Telemetries of a true anomalous event along with the autoencoder’s [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
read the original abstract

The increasing autonomy of spacecraft demands fault-detection systems that are both reliable and explainable. This work addresses eXplainable Artificial Intelligence for onboard Fault Detection, Isolation and Recovery within the Attitude and Orbit Control Subsystem by introducing a framework that enhances interpretability in neural anomaly detectors. We propose a method to derive low-dimensional, semantically annotated encodings from intermediate neural activations, called peepholes. Applied to a convolutional autoencoder, the framework produces interpretable indicators that enable the identification and localization of anomalies in reaction-wheel telemetry. Peepholes analysis further reveals bias detection and supports fault localization. The proposed framework enables the semantic characterization of detected anomalies while requiring only a marginal increase in computational resources, thus supporting its feasibility for on-board deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes a framework for eXplainable AI in onboard fault detection, isolation, and recovery for satellite attitude and orbit control subsystems. It introduces 'peepholes' as low-dimensional, semantically annotated encodings derived from intermediate activations of a convolutional autoencoder applied to reaction-wheel telemetry. The approach is claimed to produce interpretable indicators for anomaly identification, localization, and bias detection while incurring only marginal computational overhead, thereby supporting feasibility for onboard deployment.

Significance. If validated, the peephole method could offer a practical route to adding semantic interpretability to neural anomaly detectors in resource-limited space systems without prohibitive compute costs, addressing a genuine need in autonomous spacecraft operations. The work highlights an important intersection of XAI and onboard telemetry monitoring. However, the absence of any empirical results means the significance is currently prospective rather than demonstrated.

major comments (2)
  1. Abstract: The central claim that peepholes enable semantic characterization of anomalies (including bias detection and fault localization) with only marginal computational overhead is presented without any quantitative results, validation on real reaction-wheel telemetry data, error analysis, or baseline comparisons. This directly undermines assessment of the onboard feasibility argument.
  2. Abstract: The manuscript provides no details on how semantic annotations are generated for the peepholes or on the datasets and metrics used to confirm that intermediate conv-autoencoder activations yield actionable, fault-specific semantics rather than generic activation patterns. This leaves the core assumption untested and load-bearing for the interpretability benefit.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for stronger empirical grounding and methodological transparency in our proposal. The comments correctly identify that the current manuscript presents the peephole framework at a conceptual level without supporting experiments or detailed validation procedures. We will revise the manuscript to address these gaps while preserving the core contribution as a framework for interpretable onboard anomaly detection.

read point-by-point responses

  1. Referee: Abstract: The central claim that peepholes enable semantic characterization of anomalies (including bias detection and fault localization) with only marginal computational overhead is presented without any quantitative results, validation on real reaction-wheel telemetry data, error analysis, or baseline comparisons. This directly undermines assessment of the onboard feasibility argument.

    Authors: We agree that the abstract states the feasibility claims without quantitative backing, and the manuscript as submitted contains no empirical results, overhead benchmarks, or baseline comparisons. This is because the work was positioned as a conceptual framework introduction rather than a full experimental study. In the revised version we will add an experimental section reporting preliminary results on simulated reaction-wheel telemetry, including measured computational overhead, anomaly detection performance metrics, and comparisons against standard autoencoder baselines to substantiate the onboard feasibility argument. revision: yes

  2. Referee: Abstract: The manuscript provides no details on how semantic annotations are generated for the peepholes or on the datasets and metrics used to confirm that intermediate conv-autoencoder activations yield actionable, fault-specific semantics rather than generic activation patterns. This leaves the core assumption untested and load-bearing for the interpretability benefit.

    Authors: The semantic annotations are conceptually obtained by mapping low-dimensional peephole encodings to known physical fault signatures (e.g., reaction-wheel speed bias or torque anomalies) using domain knowledge of attitude control telemetry. However, the submitted manuscript indeed omits an explicit description of the annotation procedure, the datasets employed, and any quantitative metrics confirming semantic specificity. We will revise the methods section to provide a step-by-step account of the annotation generation process, specify the telemetry datasets (simulated and any available public sources), and introduce evaluation metrics that demonstrate the peepholes capture fault-specific rather than generic patterns. revision: yes

Circularity Check

0 steps flagged

No circularity: proposed peephole framework is a methodological introduction without self-referential derivations

full rationale

The manuscript introduces a framework for extracting low-dimensional peepholes from intermediate activations of a convolutional autoencoder applied to reaction-wheel telemetry, with the goal of improving semantic interpretability for anomaly detection. No equations, parameter-fitting steps, or derivation chains are present that reduce a claimed result to its own inputs by construction. The text does not invoke self-citations as load-bearing uniqueness theorems, smuggle ansatzes via prior work, or rename known empirical patterns as new predictions. Claims about marginal computational overhead and semantic characterization are presented as properties of the proposed method rather than tautological outputs, leaving the work self-contained for external empirical validation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the assumption that intermediate neural activations can be meaningfully projected into low-dimensional semantically annotated encodings without loss of critical fault information. No free parameters or invented physical entities are described. The main invented entity is the peephole concept itself.

axioms (2)
  • domain assumption Neural network activations contain extractable semantic information relevant to anomaly detection in telemetry.
    Implicit in the proposal to derive peepholes from intermediate layers.
  • domain assumption A marginal increase in computational resources is acceptable for onboard satellite deployment.
    Stated as supporting feasibility for on-board use.
invented entities (1)
  • peepholes no independent evidence
    purpose: Low-dimensional semantically annotated encodings from neural activations for interpretable anomaly detection.
    New term and method introduced to enhance explainability in the autoencoder.

pith-pipeline@v0.9.0 · 5462 in / 1419 out tokens · 55769 ms · 2026-05-10T17:37:27.618125+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

  1. [1]

    ESA Earth Observation Science Strategy

    European Space Agency (ESA). ESA Earth Observation Science Strategy. Technical report, European Space Agency (ESA), sep 2024

    work page 2024

  2. [2]

    Earth Science to Action

    National Aeronautics and Space Administration (NASA). Earth Science to Action. Technical report, National Aeronautics and Space Administration (NASA), 2024

    work page 2024

  3. [3]

    Pelican-2 & 36 SuperDoves Arrived in Vandenberg, California For Launch, December 2024

    Planet Labs PBC. Pelican-2 & 36 SuperDoves Arrived in Vandenberg, California For Launch, December 2024. Planet Pulse news article; describes Pelican-2 equipped with NVIDIA Jetson for edge AI

    work page 2024

  4. [4]

    Towards the use of artifi- cialintelligenceontheedgeinspacesystems: Challengesandopportunities

    Gianluca Furano, Gabriele Meoni, Aubrey Dunne, David Moloney, Veronique Ferlet-Cavrois, Antonis Tavoularis, Jonathan Byrne, Léonie Buckley, Mihalis Psarakis, Kay-Obbe Voss, et al. Towards the use of artifi- cialintelligenceontheedgeinspacesystems: Challengesandopportunities. IEEE Aerospace and Electronic Systems Magazine, 35(12):44–56, 2020

    work page 2020

  5. [5]

    TheΦ-Sat-1 Mission: The First On-Board Deep Neural Net- work Demonstrator for Satellite Earth Observation.IEEE Transactions on Geoscience and Remote Sensing, 60:1–1, 2022

    Gianluca Giuffrida, Luca Fanucci, Gabriele Meoni, Matej Batic, Leonie Buckley, Aubrey Dunne, Chris Van Dijk, Marco Esposito, John Hefele, Nathan Vercruyssen, Gianluca Furano, Massimiliano Pastena, and Josef Aschbacher. TheΦ-Sat-1 Mission: The First On-Board Deep Neural Net- work Demonstrator for Satellite Earth Observation.IEEE Transactions on Geoscience ...

    work page 2022

  6. [6]

    OPS-SAT Spacecraft Autonomy with TensorFlow Lite, Unsupervised Learning, and Online Machine Learning

    Georges Labreche, David Evans, Dominik Marszk, Tom Mladenov, Vasund- hara Shiradhonkar, Tanguy Soto, and Vladimir Zelenevskiy. OPS-SAT Spacecraft Autonomy with TensorFlow Lite, Unsupervised Learning, and Online Machine Learning. In2022 IEEE Aerospace Conference (AERO), volume 2022-March, pages 1–17, 2022. Cited by: 26

    work page 2022

  7. [7]

    In-orbit demonstration of a re-trainable machine learning payload for processing optical imagery.Scientific Reports, 13(1),

    Gonzalo Mateo-Garcia, Josh Veitch-Michaelis, Cormac Purcell, Nicolas Longepe, Simon Reid, Alice Anlind, Fredrik Bruhn, James Parr, and Pierre Philippe Mathieu. In-orbit demonstration of a re-trainable machine learning payload for processing optical imagery.Scientific Reports, 13(1),

    work page

  8. [8]

    Cited by: 18; All Open Access, Gold Open Access. 13

    work page

  9. [9]

    Operationalizing explainable AI in the EU regulatory ecosystem.IEEE Intelligent Systems, 2024

    Luca Nannini, Jose Maria Alonso-Moral, Alejandro Catalá, Manuel Lama, and Senén Barro. Operationalizing explainable AI in the EU regulatory ecosystem.IEEE Intelligent Systems, 2024

    work page 2024

  10. [10]

    Leveson.Engineering a Safer World: Systems Thinking Applied to Safety

    Nancy G. Leveson.Engineering a Safer World: Systems Thinking Applied to Safety. MIT Press, Cambridge, MA, 2011

    work page 2011

  11. [11]

    FDI (R) for satellites: How to deal with high availability and robustness in the space domain?International Journal of Applied Mathematics and Computer Science, 22(1):99–107, 2012

    Xavier Olive. FDI (R) for satellites: How to deal with high availability and robustness in the space domain?International Journal of Applied Mathematics and Computer Science, 22(1):99–107, 2012

    work page 2012

  12. [12]

    Control moment gyroscope anomaly detection based on sparse autoencoder

    Yuxue Zheng, Haijun Dou, Hutao Cui, and Minqiang Xu. Control moment gyroscope anomaly detection based on sparse autoencoder. InJournal of Physics: Conference Series, number 1, page 012061. IOP Publishing, 2024

    work page 2024

  13. [13]

    Anomaly detec- tion of control moment gyroscope based on working condition classification and transfer learning.Applied Sciences, 13(7):4259, 2023

    Kuan Zhang, Shuchen Wang, Saijin Wang, and Qizhi Xu. Anomaly detec- tion of control moment gyroscope based on working condition classification and transfer learning.Applied Sciences, 13(7):4259, 2023

    work page 2023

  14. [14]

    Kauffmann, Robert A

    Lukas Ruff, Jacob R. Kauffmann, Robert A. Vandermeulen, Grégoire Mon- tavon, Wojciech Samek, Marius Kloft, Thomas G. Dietterich, and Klaus- Robert Müller. A Unifying Review of Deep and Shallow Anomaly Detec- tion.Proceedings of the IEEE, 109(5):756–795, 2021

    work page 2021

  15. [15]

    Why should I trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. “Why should I trust you?” Explaining the predictions of any classifier. InProceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016

    work page 2016

  16. [16]

    Mythos of Model Interpretability.Queue, 16(3):31–57, 2018

    Zachary C Lipton. Mythos of Model Interpretability.Queue, 16(3):31–57, 2018

    work page 2018

  17. [17]

    Number NIST.IR.8312 in NIST Technical Series Publications

    Andrew Schumann.Four Principles of Explainable Artificial Intelligence. Number NIST.IR.8312 in NIST Technical Series Publications. National In- stitute of Standards and Technology, 2021

    work page 2021

  18. [18]

    Explainable anomaly detection in spacecraft telemetry.Engineering Applications of Artificial Intelligence, 133:1–15, 02 2024

    Sara Cuéllar Carrillo, Matilde Santos Peñas, Fernando Alonso, Ernesto Fabregas, and Gonzalo Farias. Explainable anomaly detection in spacecraft telemetry.Engineering Applications of Artificial Intelligence, 133:1–15, 02 2024

    work page 2024

  19. [19]

    Health indicator construction of high-speed rotating bearings in aerospace CMG based on physics-inspired machine-learning approach

    Diyin Tang, Mengtong Gong, Limei Tian, Jinsong Yu, Jiyang Zhang, and Qiang Zhang. Health indicator construction of high-speed rotating bearings in aerospace CMG based on physics-inspired machine-learning approach. IEEE Transactions on Instrumentation and Measurement, 71:1–11, 2022

    work page 2022

  20. [20]

    Innovative ml-based methods for automated on- board spacecraft anomaly detection

    Carlo Ciancarelli, Eleonora Mariotti, Francesco Corallo, Salvatore Cognetta, Livia Manovi, Alex Marchioni, Mauro Mangia, Riccardo Rovatti, and Gianluca Furano. Innovative ml-based methods for automated on- board spacecraft anomaly detection. InInternational Conference on Ap- plied Intelligence and Informatics, pages 213–228. Springer, 2022. 14

    work page 2022

  21. [21]

    Imbalanced satellite telemetry data anomaly detection model based on Bayesian LSTM.Acta Astronautica, 180:232–242, 2021

    JunfuChen, DechangPi, ZhiyuanWu, XiaodongZhao, YuePan, andQiang Zhang. Imbalanced satellite telemetry data anomaly detection model based on Bayesian LSTM.Acta Astronautica, 180:232–242, 2021

    work page 2021

  22. [22]

    Physics-informed machine learning: A comprehensive review on applications in anomaly detection and condition monitoring.Expert Systems with Applications, 255:124678, 2024

    Yuandi Wu, Brett Sicard, and Stephen Andrew Gadsden. Physics-informed machine learning: A comprehensive review on applications in anomaly detection and condition monitoring.Expert Systems with Applications, 255:124678, 2024

    work page 2024

  23. [23]

    Generalized Rank-Constrained Matrix Approximations.SIAM Journal on Matrix Analysis and Applica- tions, 29(2):656–659, 2007

    Shmuel Friedland and Anatoli Torokhti. Generalized Rank-Constrained Matrix Approximations.SIAM Journal on Matrix Analysis and Applica- tions, 29(2):656–659, 2007

    work page 2007

  24. [24]

    A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum Likelihood from Incomplete Data Via the EM Algorithm.Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1–22, 12 2018

    work page 2018

  25. [25]

    Yu han Liu and Sercan O. Arik. Explaining Deep Neural Networks using Unsupervised Clustering, 2020

    work page 2020

  26. [26]

    A General Framework for the Assessment of Detectors of Anomalies in Time Series.IEEE Transactions on Industrial Informatics, 20(10):12051–12061, 2024

    Andriy Enttsel, Silvia Onofri, Alex Marchioni, Mauro Mangia, Gianluca Setti, and Riccardo Rovatti. A General Framework for the Assessment of Detectors of Anomalies in Time Series.IEEE Transactions on Industrial Informatics, 20(10):12051–12061, 2024

    work page 2024

  27. [27]

    An introduction to ROC analysis.Pattern Recognition Let- ters, 27(8):861–874, 2006

    Tom Fawcett. An introduction to ROC analysis.Pattern Recognition Let- ters, 27(8):861–874, 2006. ROC Analysis in Pattern Recognition. 15

    work page 2006